Semi-supervised dimensionality reduction of hyperspectral imagery using pseudo-labels. (February 2018)

• Record Type: Journal Article
• Authors: Wu, Hao; Prasad, Saurabh
• Is Part Of: Pattern recognition. Volume 74(2018:Feb.)
• Page Start: 212
• Page End: 224
• Publication Date: 2018-02
• Subjects: Dimensionality reduction -- Semi-supervised learning -- Dirichlet process mixture model -- Hyperspectral data classification; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.09.003
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A new semi-supervised feature reduction approach is proposed. Pseudo-labels are generated using Dirichlet Process Mixing Model, which are then used for learning a semi-supervised dimensionality reduction projection that simultaneously preserves locality. Projection is undertaken in both raw and kernel space. Results with hyperspectral data demonstrates substantial improvement in performance. Abstract: Dimensionality reduction has been proven to be efficient in preparing high dimensional data for various tasks in machine learning. As supervised dimensionality reduction methods such as Fisher discriminant analysis (FDA) and local Fisher discriminant analysis (LFDA) tend to suffer from overfitting when only a small number of labeled samples are available, the abundant unlabeled samples could be helpful in finding a better embedding space. However, applying discriminant analysis on unlabeled data is challenging since we do not have labels for unlabeled data. In this paper, we propose a semi-supervised Semi-Supervised Local Fisher Discriminant Analysis (SSLFDA) using pseudo labels, aiming to perform discriminant analysis on both labeled and unlabeled samples. SSLFDA makes use of pseudo labels, learned from the Dirichlet process mixture model (DPMM) based clustering algorithm, to enable local Fisher discriminant analysis on unlabeled data. In addition, a kernel extension of SSLFDA is derived for non-linear dimensionality reduction. We present experimental results with real hyperspectral data to show that our method provides better classification performance compared to other existing dimensionality reduction methods.